AGENCYSCRIPT
CoursesEnterpriseBlog
đź‘‘FoundersSign inJoin Waitlist
AGENCYSCRIPT

Governed Certification Framework

The operating system for AI-enabled agency building. Certify judgment under constraint. Standards over scale. Governance over shortcuts.

Stay informed

Governance updates, certification insights, and industry standards.

Products

  • Platform
  • Certification
  • Launch Program
  • Vault
  • The Book

Certification

  • Foundation (AS-F)
  • Operator (AS-O)
  • Architect (AS-A)
  • Principal (AS-P)

Resources

  • Blog
  • Verify Credential
  • Enterprise
  • Partners
  • Pricing

Company

  • About
  • Contact
  • Careers
  • Press
© 2026 Agency Script, Inc.·
Privacy PolicyTerms of ServiceCertification AgreementSecurity

Standards over scale. Judgment over volume. Governance over shortcuts.

On This Page

Example 1: The Investor Update That Sounded Like a TweetWhat fixed itExample 2: The Support Reply That Was Too StiffWhat fixed itExample 3: The Legal Summary That Hedged Itself Into MushWhat fixed itExample 4: The Recruiting Email That Was Off-BrandWhat fixed itExample 5: The Technical Doc That Talked Down to EngineersWhat fixed itExample 6: The Multilingual Newsletter With Inconsistent PolitenessWhat fixed itWhat These Cases Have in CommonReplace adjectives with mechanicsName the reader before the toneTreat markers as dialsFrequently Asked QuestionsWhy does "write professionally" produce inconsistent results?What is the single highest-leverage register control?Should I use examples or rules to control tone?How do I stop the model from over-hedging?Does register control differ across languages?How many tone markers should I specify in one prompt?Key Takeaways
Home/Blog/Six Annotated Prompts Where Tone Either Landed or Backfired
General

Six Annotated Prompts Where Tone Either Landed or Backfired

A

Agency Script Editorial

Editorial Team

·September 8, 2019·8 min read
controlling formality and register in outputcontrolling formality and register in output examplescontrolling formality and register in output guideprompt engineering

Ask a model to "write professionally" and you will get a default register that reads like a press release written by a committee. The word "professional" means something different to a litigation partner, a Gen-Z community manager, and a hospital intake nurse. Register is not a single dial that runs from casual to formal. It is a bundle of choices about vocabulary, sentence length, hedging, contractions, jargon density, and how much distance the writer keeps from the reader. When a prompt fails to control tone, it is almost always because the instruction collapsed that bundle into one vague adjective.

The fastest way to learn register control is to look at real prompts side by side and study what changed the output. Below are six scenarios drawn from common agency and client work. Each pairs a prompt that produced the wrong tone with a revision that fixed it, plus a short note on the mechanism. The point is not to memorize phrasings but to see the underlying levers you can pull on any task.

Read these as diagnostic cases. By the end you should be able to look at a flat or mismatched output and name the specific lever that would correct it, rather than reaching again for "make it more professional."

Example 1: The Investor Update That Sounded Like a Tweet

A founder asked for a quarterly update to send to seed investors. The first prompt was simply "Write a friendly investor update about our Q3 numbers." The result opened with "Hey everyone! Big news!!" and used three exclamation points in the first paragraph. Friendly to a model with no other anchor means breezy and excited.

What fixed it

The revision specified the relationship and the reader's expectations: "Write a Q3 update for seed investors who expect candor and concision. Confident but measured. No exclamation points. Lead with the single most important number, then context." The output dropped the hype, opened with the revenue figure, and read like a competent operator talking to peers.

  • The lever was reader model, not a tone adjective. Naming who reads it constrained the register more than "friendly" or "formal" ever could.
  • Banning a specific marker (exclamation points) removed the most visible symptom of over-casualness.
  • Asking for a structural choice (lead with the number) pushed the prose toward the dry, declarative mode investors prefer.

Example 2: The Support Reply That Was Too Stiff

A SaaS team wanted warmer support responses. Their prompt said "Respond formally and professionally to this refund request." The reply was correct but icy: "We have received your inquiry regarding a refund. Per our policy, the following applies." Customers read this as cold and bureaucratic.

What fixed it

The revision moved the dial down and named the emotion to convey: "Reply to this refund request. Warm and human, but precise about the policy. Use contractions. Acknowledge the frustration in one sentence before explaining the outcome." The result said "I'm sorry the trial didn't work out — let me sort this for you," then stated the policy plainly.

  • Contractions are one of the highest-leverage register markers. Allowing or banning them shifts perceived warmth more than almost any other single instruction.
  • Naming a target emotion ("acknowledge frustration") gave the model a concrete behavior instead of an abstract register.

Example 3: The Legal Summary That Hedged Itself Into Mush

A consultant asked for a plain-language summary of a contract clause. The prompt was "Summarize this clause clearly." The model produced something so hedged — "this may, in certain circumstances, potentially apply" — that the reader could not tell what the clause did.

What fixed it

The fix targeted hedging directly: "Summarize this clause in plain English for a non-lawyer. State what it does in direct sentences. Flag genuine ambiguity only where the text is actually unclear, and say so explicitly." Hedging dropped, and the one real ambiguity got a clear callout instead of being buried in qualifiers.

  • Excessive hedging is a register failure that reads as evasive. Telling the model to reserve qualifiers for genuine ambiguity sharpens the whole piece.
  • "For a non-lawyer" set the jargon ceiling without requiring a list of banned terms.

Example 4: The Recruiting Email That Was Off-Brand

An agency with a deliberately irreverent brand voice got LinkedIn outreach that sounded like every other recruiter. The prompt named the brand but not its mechanics. For more on encoding a voice the model can reuse, see The Anatomy of a Reusable Brand Voice Prompt.

What fixed it

They added concrete voice rules: "Short sentences. One dry joke. No corporate words like 'synergy,' 'leverage,' or 'rockstar.' Talk like a smart friend who respects the reader's time." The banned-word list did most of the work, because brand voice is often defined more by what it refuses to say than by what it embraces.

  • A banned-vocabulary list is more reliable than a positive style description for protecting a distinctive voice.
  • Length constraints ("short sentences") shape register because clipped prose reads as confident and modern.

Example 5: The Technical Doc That Talked Down to Engineers

A prompt asked for "beginner-friendly" API docs, but the audience was senior engineers who found the result patronizing — every term defined, every step over-explained.

What fixed it

"Beginner-friendly" was the wrong target. The revision set the actual expertise level: "Write for engineers fluent in REST who have not used this specific API. Assume they know HTTP, auth flows, and JSON. Explain only what is unique to our service." Matching register to expertise removed the condescension. This connects to the broader practice of measuring whether tone fits the reader, covered in Scoring Whether Generated Tone Actually Fits the Reader.

  • Register failures often come from a wrong expertise estimate, not a wrong formality level.
  • "Assume they know X" is a cleaner instruction than "be technical," because it specifies the shared knowledge the prose can lean on.

Example 6: The Multilingual Newsletter With Inconsistent Politeness

A team localizing a newsletter into Japanese and German found the model defaulted to different politeness levels across languages, breaking brand consistency.

What fixed it

They anchored register per language explicitly: in Japanese, specify the politeness level (teineigo for a broad subscriber list); in German, choose Sie versus du up front. Register is language-specific, and leaving it implicit lets the model pick a default that may not match your brand. For a fuller comparison of approaches when register requirements conflict, see Choosing Between Few-Shot Examples and Explicit Tone Rules.

  • In languages with grammaticalized politeness, register is not optional styling — it is encoded in verb forms and pronouns, so it must be specified.
  • Brand consistency across locales requires a per-language register spec, not a single global instruction.

What These Cases Have in Common

Replace adjectives with mechanics

Every fix swapped a vague adjective (friendly, formal, professional) for concrete mechanics: a reader model, a banned-word list, a contraction policy, a hedging rule, an expertise assumption. The model cannot infer your intent from "professional," but it can execute "no contractions, no exclamation points, lead with the conclusion."

Name the reader before the tone

In almost every case, specifying who reads the output did more than specifying the tone. Register exists in relation to a reader. Pin the reader, and the register narrows on its own.

Treat markers as dials

Contractions, exclamation points, hedge words, sentence length, and jargon density are independent dials. Naming the specific marker you want to move is far more reliable than nudging an abstract register. A practical sequence for getting from zero to a working tone spec is laid out in Your Fastest Route to a First Reliable Tone Spec.

Frequently Asked Questions

Why does "write professionally" produce inconsistent results?

Because "professional" has no fixed meaning across contexts. To a lawyer it means cautious and precise; to a startup it means crisp and confident; to a nurse it means warm and clear. The model picks an average that rarely matches your specific context. Replace the adjective with the underlying mechanics — reader, vocabulary limits, contraction policy — and the inconsistency disappears.

What is the single highest-leverage register control?

Naming the reader, followed closely by a contraction policy. Telling the model exactly who reads the output constrains vocabulary, sentence length, and formality all at once. After that, allowing or banning contractions shifts perceived warmth more than any other single marker.

Should I use examples or rules to control tone?

Both work, but for different jobs. Rules are precise and auditable for hard constraints like banned words. Examples capture hard-to-describe voice nuances that resist explicit rules. Most strong tone prompts combine a short rule set with one or two reference samples.

How do I stop the model from over-hedging?

Tell it to reserve qualifiers for genuine ambiguity and to state conclusions directly otherwise. Over-hedging is a register failure that reads as evasive. An explicit instruction to flag real ambiguity, and only real ambiguity, sharpens the output.

Does register control differ across languages?

Significantly. Languages like Japanese, Korean, and German encode politeness in grammar, so register is not optional styling — it lives in verb forms and pronouns. Specify the politeness level per language rather than relying on a single global tone instruction.

How many tone markers should I specify in one prompt?

Usually three to five. A reader model, a contraction policy, and one or two banned-word or length rules cover most cases. Beyond that you risk over-constraining the prose into something stilted. Start small and add markers only when a specific failure appears.

Key Takeaways

  • Vague tone adjectives like "professional" and "friendly" fail because they collapse a bundle of independent choices into one undefined word.
  • Naming the reader is the highest-leverage register control; pin the reader and the tone narrows on its own.
  • Contractions, exclamation points, hedge words, sentence length, and jargon density are independent dials you can name directly.
  • Banned-vocabulary lists protect distinctive brand voices better than positive style descriptions.
  • Register failures often trace to a wrong expertise estimate, not a wrong formality level.
  • In languages with grammaticalized politeness, register must be specified explicitly because it is encoded in grammar, not styling.

Search Articles

Categories

OperationsSalesDeliveryGovernance

Popular Tags

prompt engineeringai fundamentalsai toolsthe difference between AIMLagency operationsagency growthenterprise sales

Share Article

A

Agency Script Editorial

Editorial Team

The Agency Script editorial team delivers operational insights on AI delivery, certification, and governance for modern agency operators.

Related Articles

General

Prompt Quality Decides Whether AI Earns Its Keep

Prompt quality is the single biggest variable in whether AI delivers real work or expensive noise. The model matters, the platform matters — but the prompt you write determines whether you get a first

A
Agency Script Editorial
June 1, 2026·10 min read
General

Counting the Real Cost of Every Token You Send

Tokens and context windows sit at the intersection of AI capability and operational cost—yet most business cases treat them as technical footnotes. That's a mistake that costs real money. Every time y

A
Agency Script Editorial
June 1, 2026·10 min read
General

Rolling Out AI Hallucinations Across a Team

Most teams discover AI hallucinations the hard way — a confident-sounding wrong answer makes it into a client deliverable, a legal brief, or a published report. The damage isn't just to the output; it

A
Agency Script Editorial
June 1, 2026·11 min read

Ready to certify your AI capability?

Join the professionals building governed, repeatable AI delivery systems.

Explore Certification